v0.6.8

v0.6.8 — 2026-06-11

Added

• Shim-driven reasoning configuration (#244, #245): Declarative ReasoningCapability in provider YAML
• Expanded reasoning effort ladder (#245): Six levels — minimal, low, medium, high, xhigh, max
• block_index on IR stream delta events (#246, #249)
• Model-level thinking_type in shim reasoning config (#256): Per-model overrides via model_reasoning
• Anthropic provider_metadata on tool calls, tool results, and reasoning blocks (#257)
• Response reasoning losslessness across converters (#263)
• Provider-specific reasoning field normalization (#264): MiniMax, OpenRouter, Volcengine
• Unsigned reasoning block policy (#268, #269): unsigned_reasoning_blocks: as_is | preserve in ReasoningCapability
• Argo OpenAI Chat max_tokens → max_completion_tokens transform
• Provider reasoning behavior documentation (#267): New guide/provider-reasoning.md page

Changed

• _build_ir_usage return type tightened (#253)
• Anthropic stream usage handlers deduplicated (#253)

Fixed

• Anthropic stream block index desync (#246, #249)
• Cross-provider stream block boundary synthesis (#250, #251)
• Stream usage detail propagation (#252)
• OpenAI Chat thinking.type=auto passthrough (#258)
• thinking_type=enabled fallback when budget_tokens missing
• Unsigned Anthropic reasoning blocks in Argo history (#268, #269)

See the full changelog.